How Active PU Learning Boosts Cash‑Out Fraud Detection by 3×

Team Introduction

Ant Financial Risk & Decision Center is responsible for transaction and fund risk control across domestic and overseas business scenarios, focusing on theft, fraud, marketing cheating, spam registration detection, and decision making. The team leverages big‑data mining and cutting‑edge machine learning to develop the fifth‑generation risk engine AlphaRisk, fully upgrading Ant Financial's risk control system with AI.

Background

Alpharisk marks the transition of Alipay’s risk system into the AI era. The core of the fifth‑generation engine is the AI‑driven risk identification system AI Detect , which incorporates traditional supervised algorithms such as GBDT, ensemble learning, as well as numerous deep‑learning based unsupervised feature generation methods.

Challenges in Cash‑Out Fraud Modeling

Cash‑out fraud is harder to model than theft or fraud because it lacks explicit user feedback; victims rarely report cash‑out activities, resulting in a scarcity of labeled positive samples. Consequently, most existing solutions rely on unsupervised methods (anomaly detection, graph algorithms), which have high computational costs and strict feature requirements.

Labeling Difficulties

High labeling cost: Each sample requires 5–15 minutes of expert time, limiting large‑scale annotation.

Labeling errors: Even domain experts may misclassify ambiguous cases, especially when determining negative samples.

Proposed Method: Active PU Learning

The paper introduces a method that combines Active Learning with two‑step PU Learning (referred to as Active PU Learning). This approach reduces labeling effort while improving model performance for cash‑out risk detection.

Algorithm Workflow

Algorithm: Active PU Learning
1. Generate sample pool: select required samples and assign positive labels using transferred knowledge.
2. while stopping condition not met do
3.   Sampling: select samples for labeling based on a specific sampling strategy.
4.   Labeling: manually label the selected samples.
5.   Update sample pool.
6.   Update model using two‑step PU Learning.
7. end while

Sampling Strategy

Samples are chosen based on Uncertainty & Diversity : the most uncertain samples with diverse characteristics are selected via scoring, K‑Means clustering, and mini‑batch selection, dramatically reducing time cost compared to sequential active learning.

Labeling and Sample Update

Experts label only highly confident positive samples (1). Uncertain or negative labels are placed back into the unlabeled pool (U). Positive samples are oversampled to strengthen their influence in subsequent model updates.

Model Update

The scenario follows a PU setting: multiple rounds of expert labeling expand the positive set (P), and the learner updates the model using two‑step PU Learning. GBRT (Gradient Boosting Regression Tree) serves as the base classifier, producing a final GBRT model.

Experimental Results

Three experiments validate the effectiveness of two‑step PU Learning, Active Learning, and the combined Active PU Learning framework. Using million‑scale training data, the PU‑enhanced GBRT model achieved higher accuracy (70%) than Isolation Forest (60%) and standard GBRT (≈60%). Active Learning improved model accuracy from 91% (unsupervised) to 94% (AL‑RF). The Active PU Learning model consistently outperformed or matched the baselines across percentile ranges, delivering a three‑fold increase in cash‑out detection volume at comparable accuracy.

Conclusion and Outlook

Active PU Learning effectively leverages limited labeled data and external information to build reliable fraud detection models, though it demands high‑quality manual labels and incurs greater training time than conventional GBRT. The approach has already achieved a three‑fold increase in cash‑out detection volume and is being explored for broader applications.